Traditional methods for egg fertility detection often suffer from low efficiency, high costs, and destructive sampling. While deep learning offers a non-destructive alternative, existing models struggle to balance robustness and efficiency, particularly when handling the subtle features of fertilized eggs under stringent deployment constraints. To address this, we propose a novel lightweight classification framework based on an improved YOLOv8 architecture. Our approach introduces three key innovations: a fully Ghost-based C2f (FG-C2f) module to eliminate backbone redundancy, an SCDown decoupled downsampling strategy to optimize deep-layer computation, and a minimalist depthwise convolution head for extreme parameter compression. Experimental results demonstrate that our proposed method significantly outperforms the baseline YOLOv8n-cls, achieving comparable or superior classification accuracy while reducing model parameters by over 60% and increasing inference speed by more than 40%. By achieving an optimal trade-off between precision and efficiency, this work provides a practical solution for embedded agricultural applications and offers valuable insights for designing efficient pattern recognition systems on resource-constrained edge devices.
Yue et al. (2026) studied this question.